Knowledge Base Retrieval and Recall for Lead Optimization

Data Characteristics

Lead optimization quality documents primarily include synthesis batch records, analytical test reports, stability study data, impurity profile analysis reports, and process validation files for small molecule compounds. These documents typically originate from Laboratory Information Management Systems (LIMS), Electronic Lab Notebooks (ELN), or Quality Management Systems (QMS). Data update frequency is relatively low, occurring after compound structure determination, synthesis batch completion, or during periodic stability study reports. Document structures consist mainly of structured tabular data and unstructured text descriptions. For example, synthesis steps are usually text descriptions, while analytical results often appear in tables with specific values, units (e.g., ppm, ng/mL, ℃), and detection methods. Key fields include batch number, compound ID, test item, test result, limit standard, instrument model, and operator.

Constraints on Knowledge Base Retrieval and Recall

The data characteristics of lead optimization quality documents impose several requirements on knowledge base retrieval and recall. First, the mix of structured and unstructured information in documents requires the knowledge base to jointly understand tabular data and text descriptions. This ensures accurate association of different data types during retrieval. Second, the low update frequency combined with large data volumes necessitates efficient text processing and indexing mechanisms for initial knowledge base construction and subsequent incremental updates. The precise numerical values and specific units in documents demand highly accurate recall results, preventing data misinterpretation due to semantic ambiguity. Reliance on key fields like batch number and compound ID requires the knowledge base to support exact matching and multi-field combined queries. This ensures the completeness and relevance of the recalled context. Additionally, variations in detection methods and instrument models can affect the interpretation of retrieval results, requiring sufficient background information during recall.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`800–1200 characters` (characters)	Ensures each text block contains sufficient context, covering complete experimental steps or analysis report segments, preventing critical information from being split.
`Recall count` (Recall Count)	`Top 5 entries` (top 5)	Provides an appropriate number of candidate document blocks for subsequent processing, given relevant recall. This avoids information overload while covering potential multiple related batches or test reports.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Balances recall precision and recall rate. Filters irrelevant documents, ensuring only highly relevant segments of lead optimization quality documents are returned.
`Rerank result count` (Rerank Count)	`3 entries` (3)	Reranks the initially recalled document blocks, prioritizing the quality document segments that best match the current query intent, improving user experience.
`maxContext`	`4000 tokens`	Ensures the model can fully receive and understand the content of multiple recalled document blocks during processing, especially those containing detailed experimental records and analytical data.
`UPLOAD_FILE_MAX_SIZE`	`500 MB`	Accounts for the potential upload of PDF files containing numerous charts and detailed batch records, providing sufficient single-file upload capacity.

Common Pitfalls

The text understanding model list is empty during knowledge base creation. This occurs when the corresponding text embedding model service is not correctly configured or enabled.
When calling the interface for retrieval, only 1 text block with the highest match is returned. This happens when the Recall count (Recall Count) parameter is set too low, limiting the number of returned results.
Uploading large PDF batch record files results in a long response time or processing failure. This indicates that the PARSE_FILE_TIMEOUT_SECONDS configuration is insufficient to handle the parsing time of complex documents.

Verification Steps

Upload representative lead optimization quality documents (e.g., PDFs containing synthesis batch records and analysis reports). Verify that they are parsed correctly and generate text blocks.
Use typical query statements (e.g., "stability data for compound A batch B," "limit standard for impurity C") for retrieval. Check if the recalled text blocks accurately contain the information snippets required by the query.
After multiple retrievals, examine the distribution of similarity scores in the recall results. Adjust the Similarity threshold (Similarity Threshold) based on actual business needs to ensure recall results are neither too broad nor miss critical information.
Simulate user queries in different scenarios to verify the knowledge base's recall effectiveness across various query types (e.g., exact queries, fuzzy queries, multi-field combined queries).

Note: The values provided are common starting points. Measure them against your own samples for optimal configuration.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.